Revert conv change which did not actually help

Author: Dobiasd

include/fdeep/convolution.hpp

@@ -139,29 +139,21 @@ namespace internal {
 
             const auto input = get_im2col_mapping(in, f_width, f_depth, 1, mapping_width, 0, y_filt);
 
-                Eigen::Map<Eigen::Matrix<float_type, Eigen::Dynamic, Eigen::Dynamic>, Eigen::Unaligned>
+            Eigen::Map<Eigen::Matrix<float_type, Eigen::Dynamic, Eigen::Dynamic>, Eigen::Unaligned>
                 output_temp_map(&output_temp.get_ref_ignore_rank(tensor_pos(0, 0, 0, 0, 0)),
                     static_cast<EigenIndex>(out_depth),
                     static_cast<EigenIndex>(mapping_width));
 
-            // Use = (beta=0) for the first filter row to avoid reading the
-            // zero-initialized output_temp values, saving memory bandwidth.
-            if (y_filt == 0) {
-                output_temp_map.noalias() = filter * input;
-            } else {
-                output_temp_map.noalias() += filter * input;
-            }
+            output_temp_map.noalias() += filter * input;
         }
 
-        // Add the valid portion of each output_temp row into the bias-initialized
-        // output tensor. Process one full row (out_width * out_depth contiguous
-        // elements) per iteration using Eigen's SIMD-vectorized addition.
+        // Dropping the superfluous results from "between" the rows.
         for (std::size_t y_out = 0; y_out < out_height; ++y_out) {
-            const float_type* src = &output_temp.get_ref_ignore_rank(tensor_pos(0, 0, y_out, 0, 0));
-            float_type* dst = &output.get_ref_ignore_rank(tensor_pos(0, 0, y_out, 0, 0));
-            const EigenIndex n = static_cast<EigenIndex>(out_width * out_depth);
-            Eigen::Map<Eigen::Array<float_type, Eigen::Dynamic, 1>>(dst, n) +=
-                Eigen::Map<const Eigen::Array<float_type, Eigen::Dynamic, 1>>(src, n);
+            for (std::size_t x_out = 0; x_out < out_width; ++x_out) {
+                for (std::size_t z_out = 0; z_out < out_depth; ++z_out) {
+                    output.get_ref_ignore_rank(tensor_pos(0, 0, y_out, x_out, z_out)) += output_temp.get_ref_ignore_rank(tensor_pos(0, 0, y_out, x_out, z_out));
+                }
+            }
         }
 
         return output;